Han Wei Tan
Most "AI made a discovery" claims collapse the moment someone asks for a control.
Core trick
Mines co-essentiality across two CRISPR screens, then requires independent physical-interaction evidence and a confound filter before accepting any predicted complex member.
Tags
Description
Most "AI made a discovery" claims collapse the moment someone asks for a control. I built the opposite: a self-falsifying engine that predicts the missing members of protein complexes from CRISPR genetics and refuses to trust itself.
The engine mines co-essentiality across two independent genome-wide CRISPR screens (Broad DepMap, 1,100 cell lines; Sanger Project Score, 324), then forces every prediction through independent physical-interaction evidence (BioPlex, hu.MAP, CFdb) and a confound firewall that rejects genetic-context artifacts. Benchmarked honestly: leave-one-out recall@10 is about 28% (roughly 475x chance), AUROC about 0.69 in both screens, and it independently re-derives dark-gene functions only published in 2022 to 2025.
Why it matters, shown in a live test: I asked Claude Opus 4.8 with no rules for a protein's complex partners. It confidently returned ten, and its built-in reviewer approved all ten. Under the engine's rules, only one clears every check. The honest breakdown is four clear false positives, five real members held back as unproven, and one you can fully trust. Same model in both halves, so the rules, not a smarter AI, earn the trust.
On Gladstone's Krogan head and neck cancer interactome, a hand-seeded KEAP1-CUL3-NRF2 pathway recovers UBE2M as a positive control (known biology, labeled as such; the automated pipeline's top CUL3 hit is ARIH1, and I say so). I then reproduced the work in Claude Science, whose independent reviewer caught me calling a correlation "replicated" when it sat below my own 0.3 threshold. The safety net fired on my own analysis.
It is a reusable rigor layer, not a one-off result: a scientist can run `python scripts/complete.py --seed BRCA1,BRCA2,PALB2` and get BARD1 (BRCA1's obligate partner) as the sole trustworthy hit, with every candidate labeled PASS, WEAK, or NO-PHYSICAL. The repo also ships an honest map of ten attempts, most of which collapsed under proper controls, because a negative established rigorously beats a positive established loosely. Built by Claude Code, audited by Claude Science. Not a hype reel. A trust layer.
README excerpt
# The Honest Discovery Engine ### A self-falsifying engine for finding missing protein-complex members in a cancer interactome **Built with Claude: Life Sciences (Gladstone Institutes × Anthropic × Cerebral Valley). Research track.** MIT licensed. *This is the project overview. For the engine internals, how to run it on your own complex, and the full honest map, see [`har_discovery/README.md`](har_discovery/README.md).* > Most "AI made a discovery" claims die the moment someone asks for a control. I built the > opposite. My engine predicts the **missing members of protein complexes** from genetics, forces > every prediction through **independent physical-interaction evidence**, and **flags its own false > positives**. I then reproduced it inside **Claude Science**, whose reviewer audited my analysis > against its own execution log and **caught my headline number asserted instead of computed, then > made me prove it.** Pointed at a Gladstone dataset (the Krogan head & neck cancer interactome), it > recovers known cancer biology, rejects confounds, and surfaces testable missing-member hypotheses, > each with an honest confidence rather than a hype headline. ## The one idea A physical pull-down (AP-MS) misses transient or condition-dependent members of a protein complex. Genetics sees them anyway: genes that are **essential-together** across two independent CRISPR screens (**Broad DepMap, 1,100 cell lines**, and **Sanger Project Score, 324 cell lines**) tend to work togeth